Avaliação de uso do coeficientes mel-cepstrais na representação das características vocais de um locutor.

Abstract

Biometric identification of individuals has been widely used as a security mechanism for accessing computer systems or restricted environments. Biometric systems have been developed to perform identification through fingerprint, iris, or voice, for example. Using the voice as a biometric identifier has been increasingly possible due to significant advances in digital processing of speech signals area. This research aims to evaluate the efficiency of mel-frequency cepstral coefficients in the representation of the characteristics of a speaker in an automatic speaker verification. The techniques used to construct the automatic speaker verification system aiming at a hardware implementation included the use of: (i) melfrequency cepstral coefficients, like feature vector; (ii) vector quantization, in patterning modelling; and (iii) a decision rule, based on Euclidean distance. The system used for evaluation in the representation of the characteristics of a speaker is a modification of another automatic speaker verification system using linear predictive coding coefficients for the representation of the vocal characteristics of a speaker. It was implemented using C++ for the training phase, and SystemVerilog for the verification phase. The results using mel-frequency cepstral coefficients were 99.34% in the hit rate, 0.17% to error rate and 0.49% to unknown response rate, compared respectively to 96.52% in success rate, 0.90% to error rate and 2.58% to unknown rate using the linear predictive coding coefficients.